Social Media and the News Industry

Transcription

1 Social Media and the News Industry Alexandre de Cornière and Miklos Sarvary October 2, 2017 Abstract The growing influence of internet platforms acting as content aggregators is one of the most important challenges facing the media industry. We develop a parsimonious model to understand the impact of content bundling by a social platform. In our model consumers can access news either directly through a newspaper s website, or indirectly through a platform, which also offers social content. Even though the platform shares revenues with newspapers whose content it publishes, content bundling harms newspapers. Its effect on news quality and news consumption depends on the media market structure and on whether the platform can personalize the content bundle. Keywords: User-Generated Content (UGC), Media Competition, News Quality Alexandre de Cornière (alexandre.de-corniere@tse-fr.eu) is Assistant Professor of Economics at Toulouse School of Economics, Université Toulouse I Capitole, and Miklos Sarvary (miklos.sarvary@columbia.edu) is the Carson Family Professor of Business at Columbia Business School. We thank Charles Angelucci, Alessandro Bonatti, Markus Reisinger, Martin Peitz, Paul Seabright, Anna D Annunzio, participants of the 4th Annual BU Platform Symposium and of the 2017 Petralia Sottana Workshop for helpful comments and discussions. We are grateful to the Net Institute ( for financial support. All errors are ours. 1

2 1 Introduction With hundreds of millions of daily active users, a few large social networks have become the dominant online media outlets for most people. The largest among these, Facebook has reached over two billion active members across the globe who, on average, spend about an hour each day on the platform. Other successful social platforms include Tencent s WeChat in China and VKontakte in Russia. If, in their early days, these platforms were mostly used as a way for users to share personal stories and pictures, their role has progressively evolved into one of content aggregation: an important share of the content displayed on their websites is produced by third party publishers, who use the platforms as an alternative to their own website to reach consumers. The news industry in particular has been affected by this change: studies show that more than 50% of consumers use social media as a source of news, and 14% as their main source (Gottfried and Shearer, 2016; Mitchell et al., 2017; Reuters, 2016). Facebook has recently surpassed Google as the main external source of traffic to newspapers websites (Alpert, 2015; Constine, 2016). 1 This situation is a double-edged sword for publishers: social platforms provide the opportunity to reach a wider audience, yet newspapers worry about the growing power of platforms, for fear of losing their privileged relationship with readers, and eventually most of their revenues (Thompson, 2017). Such content bundling is not merely a result of consumers actively sharing news stories through the platform, but is a deliberate strategic choice by the platforms. Indeed, even though the content that a consumer gets exposed to depends on the behavior of his friends, platforms retain a considerable amount of control over the content that is displayed, and can choose to emphasize one type of content over another. 2 The goal of this paper is to study how the emergence of social platforms as a major source of news affects newspapers incentives to provide quality content, their profits, as well as consumers news consumption. In our basic framework, a social platform and a newspaper (or publisher), both advertising-supported, compete for consumers attention. The newspaper produces news stories and maintains a website which only offers news content. The social platform relies on its users to produce user-generated content (UGC), such as personal stories or pictures. On its website, alongside UGC, the platform can also 1 For some news providers, Facebook s dominance is even more pronounced. For example, Buzzfeed, a leading online publisher valued at close to $1.5 billion derives 75% of its traffic from Facebook. 2 The editorial role of platforms has been a recent object of controversy, with some critics demanding that dominant platforms be held responsible for the content displayed on their websites. 2

3 show news stories produced by the publisher. We refer to the strategy of showing a mix of news and UGC as content bundling. Under content bundling, the platform gives a share of its revenues to the publisher. Consumers have limited attention, and are heterogeneous in their demand for news, that is in the share of their attention they would like to devote to news content. The demand for news depends on its quality, which is the result of an investment by the newspaper. Consumers can freely allocate their attention across the two websites, but, when on the platform, have to consume the bundle that is offered to them. We show that, provided the share of consumers who want to see no news is not too large, some content bundling occurs in equilibrium. Indeed, it allows the platform to host some of consumers news consumption and some of the associated revenues. To understand the effects of content bundling, we compare it to a benchmark in which the platform only displays UGC. We find that the newspaper s profit is lower under content bundling. Even though the platform showing news is not necessarily bad for the newspaper, the latter is harmed by content bundling because the platform strategically chooses the mix of content it offers so as to maximize its revenue, which comes at the expense of the newspaper s revenue. Content bundling also causes news consumption to be distorted upwards. Whether quality increases or decreases depends on the relative magnitude of a softening effect (negative - the newspaper s gain to attracting traffic is lower if it gets a share of the platform s revenues) and a composition effect (ambiguous - traffic may become more or less sensitive to quality under content bundling). If the newspaper can prevent bundling by opting-out, then content bundling leaves it indifferent but results in unambiguously lower investment in quality. Finally, when the platform has the ability to customize the bundle at an individual level, it can monopolize consumer s attention. News quality is then lower than without content bundling. We then extend our framework to allow for more than one publisher on the market. There again newspapers are harmed by content bundling. However the effect of bundling on quality depends on the market structure: when newspapers are local monopolists, content bundling leads to an increase in quality, whereas when newspapers compete head to head, quality goes down under content bundling. The paper is organized as follows. In Section 2, we review the relevant literature. In Section 3 we present the model. We solve both the benchmark case and the case under content bundling in Section 4. In Section 5 we present three extensions of our framework 3

4 with a single newspaper: customization, consumer single-homing, and opt-out by the newspaper. In Section 6 we present our two models of newspaper competition. Section 7 concludes. 2 Relevant literature The paper is related to a number of literature streams, first, and foremost to the broadening literature on news/media aggregators (see Peitz and Reisinger (2015) for an extensive summary on this literature). As in our paper, the central question is how these intermediaries impact the consumption of news as well as the quality of content produced. For example, Jeon and Nasr (2016) consider a model with horizontal and vertical differentiation between newspapers that choose the quality of news on multiple issues. The aggregator is modeled as an exogenous process that helps consumers discover news items of high quality. They identify the main trade-off faced by newspapers, namely the relative impact of the aggregator s business-stealing effect and demand-expansion effect, and predict that the presence of an aggregator leads newspapers to increase their quality investments. In earlier work Dellarocas, Katona, and Rand (2013) come to a similar conclusion using a model where the aggregator emerges endogenously in a model where sites can link to each others content. 3 Importantly, this body of research argues that the quality of news is likely to increase as the presence of aggregators provides a strong incentive to (over)invest in quality. Our model suggests that content bundling by a social platform has a very different impact from that of news aggregators. A recent group of empirical papers examine the impact of aggregators on the news industry. Using disputes between Google News and Spanish publishers (Athey, Mobius, and Pal (2017), Calzada and Gil (2016)) or the Associated Press (Chiou and Tucker (2015)), empirical research finds that Google News increases overall news consumption. In particular, Athey, Mobius, and Pal (2017) document that this effect is mostly present for small publishers, who cannot rely on brand recognition to attract users and therefore benefit most from the aggregator. In relation to the theoretical work on aggregators, these papers suggest that the demand-expansion effect of aggregators dominates. We find a similar effect: the social platform increases the aggregate consumption of news. However, we find that publishers are always worse off as most of the increased consumption of news is intermediated by the platform, and therefore less profitable than direct traffic. Our work specifically focuses on social networks as news intermediaries, the major 3 See also Rutt (2011) who focuses on newspapers revenue models and finds that increased competition among news providers increases the advertising-based (as opposed to subscription-based) revenue model and also leads to higher equilibrium content quality. 4

5 difference being that these platforms also host user-generated content (UGC) that directly competes with the content of publishers (see Luca (2015) for a summary of the economics literature on UGC). This is relevant because, increasingly, it is such platforms (as opposed to search engines) that generate traffic to news content. Yildirim, Gal-Or, and Geylani (2013) study the effect of UGC on the horizontal competition between news providers, but they do not consider the presence of an endogenous intermediary as we do. Theoretical research on UGC and social networks specifically is scarce and focuses mostly on network formation. 4 Our model assumes a monopolistic platform with a given audience that decides on what content mix to serve to its members. Finally, our framework assumes multi-homing but we abstract away from the core concern of the multi-homing literature applied to media, namely that it may lead to inefficient (duplicate) advertising when an advertiser is present on multiple publishers (see, Ambrus, Calvano, and Reisinger (2014), Athey, Calvano, and Gans (2017), and Anderson, Foros, and Kind (2016) for a detailed treatment of this issue). As Alaoui and Germano (2016), we also assume that consumers are time constrained in their consumption of media and our results resonate to theirs in that competition between content suppliers (including the social network) distort consumers media consumption. However, we focus on consumers time allocation across qualitatively different content providers and we abstract away from the editorial process of publishers when multiple topics are present. Interestingly, in our framework, multi-homing does not result in increased competition (or less differentiation) and actually softens competition in many cases. This softening of competition contributes to the decrease in the equilibrium quality of news. 3 The model We consider a model where consumers can consume two kinds of content: news and user-generated content (UGC). News stories are produced by a monopolist newspaper (indexed by 1), who must invest c(q) to achieve a quality q, where c is increasing and convex. User-generated content is produced by users of a monopolist social platform (indexed by 0), at no cost for the platform. UGC quality is exogenous. Consumers have heterogeneous preferences regarding content. A consumer of type θ who consumes a quantity x of news (of quality q) and y of UGC derives a utility U(x, y, q, θ), non-decreasing in x and y. We assume that U x,θ 0, 5 i.e. that high types have a larger marginal utility for news content. News quality increases the marginal 4 See, for example Bala and Goyal (2000) and Jackson and Wolinsky (1996) for earlier models, and Jackson (2010) for a review. See also Zhang and Sarvary (2015) who consider local network effects. 5 U x,θ is the cross derivative of U with respect to x and θ. 5

6 utility of news consumption: U x,q > 0. However, this effect is weaker for higher levels of quality: U x,q,q 0. 6 We assume that θ is distributed according to a continuous c.d.f. F, of density f, on a support [θ, θ]. For simplicity we assume that the distribution of types has no atoms. For now, we assume that θ is a consumer s private information. 7 Consumers have an attention constraint: x + y 1. For a given quality q, a type θ consumer s demand for news ˆx(θ, q) is the solution to max U(x, y, q, θ) s.t x 0, y 0 and x + y 1. x,y From our assumptions, ˆx(θ, q) is non-decreasing in both its arguments. Moreover, we assume that ˆx(θ, q) > 0 for all θ > θ, and that ˆx(θ, q) q,q 0. Similarly, ŷ(θ, q) is the demand for UGC. We assume that consumers have no outside option, so that the attention constraint is always binding and ŷ(θ, q) = 1 ˆx(θ, q). Example: For the sake of illustration, we sometimes use the following utility function: U(x, y, q, θ) = (α(q + θ) + βθq) ln(x) + y (1) The associated demand for news is ˆx(θ, q) = max{min{α(q + θ) + βθq, 1}, 0}. 8 We then assume that α 0, and that the parameters of the model are such that ˆx ˆx 0 and 0. θ q When β > 0, high types demand for news is more sensitive to quality than low type s demand. We refer to this as the linear model. Special cases include (α, β) = (1, 0) (what we refer to as the additive model, particularly convenient to obtain closed form solutions) and (α, β) = (0, 1) (the multiplicative model). Even though consumers have preferences over contents, they cannot directly choose which content they consume. Instead, they allocate their unit of attention across two websites: one operated by the newspaper, and one by the platform. While the newspaper s website can only offer news content, the key feature of our model is the platform s ability to display news from the newspaper alongside its own UGC. Such content bundling is a strategic choice: the platform decides the share λ of news that consumers are exposed to when they visit its website. If a consumer spends t 0 units of time on the platform s website, he therefore consumes a quantity t 0 (1 λ) of UGC, and a quantity t 0 λ of news 6 U x,q,q is the third-order partial derivative. At this point, we impose no restriction on the sign of U x,q,θ, that is, we do not specify whether high types or low types demand for news is more sensitive to quality. 7 Later, when we consider the personalization of the newsfeed by the platform, we explore the implications of relaxing this assumption. 8 In the rest of the paper we drop the max and min operators to ease notations, but demand for news and for UGC should always be thought of as being between 0 and 1. 6

7 (on top of the news he gets directly from the newspaper s website). Websites are purely advertising-supported. We normalize the monetary value of one unit of attention by a consumer to one. 9 Thus, when a consumer spends t 1 units of time on the newspaper s website (what we call direct traffic), the newspaper generates direct revenues of t 1. The newspaper also derives revenues from indirect traffic, i.e. from the news stories that consumers are exposed to while on the platform s website. More specifically, we assume that if the platform shows a share λ of news and if a consumer spends t 0 units of time on its website, the newspaper s indirect revenue is t 0 λ(1 φ), where φ [0, 1] is the share of news-related ad revenues that the platform keeps for itself. The platform s revenue is then t 0 (1 λ + λφ). One can interpret advertising revenue sharing between the platform and the newspaper either as explicit payments between them, or as capturing the idea that the consumer divides his attention between the two. In the baseline model, the revenue sharing rule (φ, 1 φ) is exogenous, but we relax this assumption in Section 5.3. Timing and equilibrium: The timing is as follows: at τ = 1, the newspaper chooses a quality q, publicly observed, and incurs the cost c(q). We view q as a long-term strategic choice. At τ = 2, the platform chooses the share of news λ it shows to its users. At τ = 3, consumers observe λ and choose t(θ, q, λ), the time they spend on the platform as a function of their type, of the quality of news and of the platform s content mix. We look for subgame perfect equilibria. 4 Equilibrium analysis 4.1 Benchmark: UGC-only newsfeed As a benchmark, we start with the case in which the platform cannot bundle news content alongside UGC (i.e. λ = 0). After observing q, consumers choose how much attention to allocate to the platform and to the newspaper. Because the platform only offers UGC, and there are no cost associated to switching from one media to the next, consumers can consume their desired basket of content. A consumer of type θ then spends ˆx(θ, q) on the newspaper site, and 9 A priori, we have no reason to assume that either website is more efficient at advertising. We discuss this assumption below. 7

8 1 ˆx(θ, q) on the platform. The total time spent on the newspaper s website is therefore T 1 (q, λ) λ=0 = θ θ ˆx(θ, q)df (θ). (2) Profit is concave in q, and the optimal quality for the newspaper, denoted q, is the solution to max q T 1 (q, 0) c(q), that is q solves: T 1 ( q, 0) q = c ( q). (3) Let π 1 be the newspaper s associated profit. We now turn to the analysis of the game where the platform can freely choose λ, and proceed by backward induction. 4.2 Consumers: allocation of attention with content bundling At τ = 3, if news quality is q, a consumer of type θ would like to consume a quantity ˆx(θ, q) of news. By spending t units of time on the platform, and 1 t on the newspaper, he gets a quantity of news, x(t, λ) = tλ + (1 t) and a quantity of UGC, y(t, λ) = t(1 λ). If λ ˆx(θ, q), the consumer s demand for news is more than satisfied by the platform alone. Such a consumer then decides to spend all his time on the platform, t(θ, q, λ) = 1. Consumers such that 1 > ˆx(θ, q) > λ can achieve their optimal content mix by spending t(θ, q, λ) on the platform such that t(θ, q, λ)(1 λ) = ŷ(θ, q) t(θ, q, λ) = ŷ(θ, q) 1 λ. Finally, if ˆx(θ, q) = 1, the consumer allocates all his attention to the newspaper, that is t(θ, q, λ) = 0. We denote by ˆθ 1 (q, λ) the solution to ˆx(θ, q) = λ, i.e. the largest type who does not visit the newspaper, and by ˆθ 2 (q) the smallest solution to ˆx(θ, q) = 1, i.e. the lowest type who does not visit the platform. We sometimes omit the arguments and simply write ˆθ 1 and ˆθ 2. Summarizing, we obtain Lemma 1 below. Lemma 1. (Optimal allocation of attention) When the newspaper is of quality q and the platform shows a share λ of news content, a consumer of type θ allocates a share t(θ, q, λ) of his attention to the platform, where t(θ, q, λ) = 1 if θ ˆθ 1, t(θ, q, λ) = 1 ˆx(θ,q) 1 λ if θ (ˆθ1, ˆθ 2 ), 8

9 1 t(θ, q, λ) 1 λ direct traffic to newspaper demand for news news consumption ˆθ 1 (q, λ) ˆθ2 (q) θ Figure 1: Individual demand for news and direct traffic to newspaper under content bundling t(θ, q, λ) = 0 if θ ˆθ 2. In the benchmark where λ = 0, consumers allocate a share ˆx(θ, q) of their attention to the newspaper. When λ > 0, that share is lower, because part of the demand for news is already satisfied by visiting the platform. More generally, any increase in λ shifts attention from the newspaper to the platform, a point we elaborate on when we discuss the choice of λ. While this effect does not directly affect consumers such that θ > ˆθ 1, whose consumption of news is still ˆx(θ, q), it introduces a consumption distortion on lower types, who, even though they stop visiting the newspaper, end up consuming too much news relative to what they would like (λ > ˆx(θ, q)). 10 Figure 1 illustrates the link between demand for news and news consumption under content bundling. Note also that, for a given quality q, total news consumption increases with λ. Indeed, a consumer s news consumption is max{ˆx(θ, q), λ}. Whether equilibrium news consumption 10 Too much news does not mean that consumers are forced to consume news content that brings them negative utility. They enjoy the news content, but would prefer UGC instead. 9

10 increases is more ambiguous, as it will depend on the effect of λ on q, an effect we examine below. 4.3 Platform: optimal content bundling Suppose that news quality is q. If the platform displays a share λ of news content, the total amount of attention that it receives is T 0 (q, λ) = θ θ t(θ, q, λ)df (θ). (4) Each unit of attention generates a revenue (1 λ + λφ), so that the platform s profit is π 0 (q, λ) = (1 λ + λφ)t 0 (q, λ). The platform s trade-off is the following: by showing more news content (increasing λ), the platform can receive more of the consumers attention, by the logic discussed in the previous subsection. However, showing more news leads to lower advertising revenue per-unit of attention. The next proposition is our first main result: Proposition 1. The platform always shows some news content: λ > 0. Proof. The derivative of the platform s profit with respect to λ is π 0 (q, λ) λ = (1 φ)t 0 (q, λ) + (1 λ(1 φ)) T 0(q, λ). (5) λ Using Lemma 1 and equation (4), we have T 0 (q, λ) λ = ˆθ 1 λ f(ˆθ 1 ) ˆθ 1 λ 1 ˆx(ˆθ 1, q) 1 λ } {{ } =1 f(ˆθ 1 ) + ˆθ2 1 ˆx(θ, q) df (θ). (6) ˆθ 1 (1 λ) 2 Evaluating the derivative of the profit at λ = 0, we thus get π 0 (q, 0) λ = φt 0 (q, 0) F (ˆθ1 (q, 0)). (7) By our assumption that ˆx(θ, q) > 0 for all θ > θ, we have ˆθ 1 (q, 0) = θ. Because F is atomless, we thus obtain π 0(q,0) = φt (q, 0) > 0. This proves the result. λ The intuition for Proposition 1 is the following. When λ = 0, consumers get all their news from the publisher. By slightly increasing λ, the platform displaces part of this news consumption towards itself, thereby increasing the attention it receives, attention valued at φ. The potential cost of doing so is to show news to some users who would have preferred to see UGC. Because λ is very small, this effect is negligible. 10

11 The result is sensitive to our assumption that the mass of consumers who want to consume no news is zero. Indeed, if such a mass was large enough, the cost of serving news instead of UGC to these news drop-outs could overcome the benefit from increased attention by news consumers. However, note that the result would also holds if the mass of consumers whose demand for news is zero was small enough. Examples: To better understand some of the forces that determine the optimal choice of λ, we use the linear model (see Equation 1) assuming that the distribution of types is uniform on [0, 1]. Using Lemma 1 to obtain individual demands, and integrating these demands over the set of types, we find that the total time spent on the platform is: 11 It follows that λ (q) = αq + φ 2(1 φ) T 0 (λ, q) = 1 + λ 2αq 2(α + βq). (8). The first term (αq) corresponds to the demand for news of the lowest type (θ = 0). Because all consumers want to consume at least a quantity αq of news, the platform must choose λ αq. Beyond this quantity, the platform s optimal strategy depends on the share φ of revenues it captures when it shows news: for large values of φ the platform has an incentive to show a lot of news content to its users. 4.4 Newspaper: choice of quality Besides understanding the strategic incentives of the platform to provide news content to its users, we seek to assess the effects of content bundling on the news industry, i.e. on newspaper s profit and choice of quality. The newspaper s profit is π 1 (q, λ) = T 1 (q, λ) + (1 φ)λt 0 (q, λ) c(q) R 1 (q, λ) c(q), where R 1 (q, λ) denotes the newspaper s advertising revenues. We assume that the primitives are such that this profit is concave in q. 12 Similarly, define R 0 (q, λ) (1 φ(1 λ))t 0 (q, λ), which represents the platform s revenues. Notice that R 0 (q, λ) + R 1 (q, λ) = T 0 (q, λ) + T 1 (q, λ) = 1 for any (q, λ). In period τ = 1, acting as a Stackelberg leader, the newspaper knows that the platform will choose λ = λ (q). Its objective function is thus π 1 (q, λ (q)) = R 1 (q, λ (q)) c(q) = 1 R 0 (q, λ (q)) c(q). Because λ (q) maximizes R 0 (q, λ), the envelope theorem implies that dπ 1(q,λ (q)) dq = π 1(q,λ (q)) q. 11 Here we only write T 0 for cases where λ αq, which is always true in equilibrium. 12 Concavity holds for instance in the linear model with uniform distribution of types. 11

12 Using the notation λ = λ (q ), the newspaper s first-order condition then writes (1 (1 φ)λ ) T 1(q, λ ) q = c (q ). (9) Comparing (3) and (9), one can distinguish two effects of content bundling by the platform: a softening and a composition effect. The softening effect corresponds to the smaller return to a marginal increase in direct traffic T 1, from 1 (in the benchmark) to 1 (1 φ)λ. When the platform bundles content, the newspaper collects a share of its revenues, and increasing T 1 is less valuable. The softening effect reduces the incentives to invest under content bundling. The composition effect works as follows: under content bundling, direct traffic to the publisher T 1 (q, λ ) only comes from consumers such that ˆx(θ, q) > λ, whereas in the benchmark direct traffic T 1 (q, 0) comes from all consumers. The number of consumers who adjust their viewing pattern following an increase in q is thus smaller under content bundling. However, because the time spent on the newspaper s website by these consumers is 1 t (θ, q, λ ) = ˆx(θ,q) 1 λ, it is more responsive to an increase in q than under the benchmark, where 1 t (θ, q, 0) = ˆx(θ, q). The overall sign of the composition effect, and therefore the effect of content bundling on news quality, is ambiguous in general. We can say more when we adopt a linear specification: Proposition 2. Suppose that ˆx(θ, q) = α(θ + q) + βθq, and that θ is uniformly distributed over [0, 1]. Then: (i) If β 0, quality goes down under content bundling compared to the benchmark. (ii) If β > 0 and φ 1, quality is higher under content bundling. Proof. When ˆx(θ, q) = α(θ + q) + βθq, the total time spent on the newspaper s website is T 1 (q, λ) = 2(α+q(α+β)) (1+λ), and therefore 2 T 1 (q,λ) β If β 0, we have 2 T 1 (q,λ) q λ 2(α+βq) q λ = 2(α+βq) 2. 0, which implies that direct traffic to the newspaper is less sensitive to quality under content bundling compared to the benchmark. This means that both effects (softening and composition) go in the same direction, and quality is unambiguously lower when the platform bundles content than under the benchmark. If β > 0 then T 1 becomes more sensitive to increases in q under content bundling, and so the two effects go in opposite direction. When φ is large enough (e.g. φ 1) then the softening effect vanishes (as the newspaper gets no revenue from indirect traffic), and only the composition effect remains, which leads to a higher quality than under the benchmark. Intuitively, under content bundling, low types consume news mostly from the platform. Direct traffic to the newspaper (T 1 ) is therefore mostly composed of high types. When β < 0, direct traffic is thus less sensitive to q than under the benchmark, whereas the opposite holds when β > 0. When the sensitivity of direct traffic is higher under content 12

13 bundling (i.e. when β > 0), the effect also need to be strong enough so as to offset the softening effect. When φ is close to 1 the softening effect is small: the newspaper does not value indirect traffic enough for competition to be relaxed. Newspaper profits While in equilibrium news quality may increase or decrease, we find that the newspaper s profit unambiguously declines with content bundling by the platform: Proposition 3. The newspaper s profit is lower under content bundling than under the benchmark. Proof. Because λ is chosen optimally by the platform, we have, for any q, R 0 (q, λ (q)) > R 0 (q, 0). This is true in particular for q = q : R 0 (q, λ ) > R 0 (q, 0). Since R 0 (q, λ) + R 1 (q, λ) = 1, the previous inequality rewrites R 1 (q, λ ) < R 1 (q, 0). Substracting c(q ) from each side, we get π 1 (q, λ ) < π 1 (q, 0). By revealed preferences, we know that π 1 (q, 0) π 1 ( q, 0), which implies that π 1 ( q, 0) > π 1 (q, λ ). Even though content bundling by the platform may soften competition and increase total news consumption, it cannot benefit the newspaper. The reason is that λ is chosen optimally by the platform to increase its revenue, which mechanically reduces the newspaper s revenue. The potential saving on costs is never enough to compensate this loss. 5 Extensions This section explores three extensions to the basic model, still assuming a monopolist newspaper. In the first, we allow the platform to offer personalized content to each of its users. In the second we assume that consumers face large switching costs and cannot multihome. In the third, we allow the newspaper to remove its content from the platform, and look at a bargaining game between the two firms. 5.1 Personalized newsfeed In the previous analysis, the platform does not have the ability to customize the mix of content it offers to each consumer. In practice however, a firm like Facebook offers different mixes to different users, leveraging the considerable amount of data it has gathered about them. We now introduce personalization to our model by assuming that the platform can observe consumers types and can condition λ on both q and θ. The timing is thus as follows: at τ = 1, the newspaper chooses q. At τ = 2 the platform observes q and θ, and chooses λ(θ, q). At τ = 3, consumers optimally allocate 13

14 their attention between the newspaper and the platform. Let ˆq be the equilibrium quality in this case. We have the following proposition: Proposition 4. When the platform can personalize the newsfeed: 1. The platform chooses λ(θ, q) = ˆx(θ, q). 2. Consumers allocate all their attention to the platform. 3. News quality is lower than under the benchmark: ˆq < q. Proof. Given θ and q, the platform clearly wants to offer λ(θ, q) = ˆx(θ, q): showing less news would induce the consumer to allocate some of his attention to the newspaper, while consuming the same amount of UGC. Showing more news would not increase the time spent on the platform, but would reduce the profitability of this time. Consumers then find it optimal to allocate all their attention to the platform. To see that quality is lower than under the benchmark, note that, for a given q, the overall consumption of news is λ(θ, q)df (θ) = ˆx(θ, q)df (θ) = T 1 (q, 0) (see expression (2)). Because all the news consumption occurs on the platform, the newspaper s profit is then (1 φ)t 1 (q, 0) c(q), instead of T 1 (q, 0) c(q) under the benchmark. The marginal return to investment is then lower than under the benchmark. Perfect personalization allows the platform to monopolize consumers attention, turning the newspaper purely into a content supplier to the platform. For a given quality level q, consumers utility is maximized. However, by lowering the returns to investment compared to the benchmark, this leads to a decreased quality of news. The assumption of perfect personalization allows us to get a clean result, but is a strong one. An alternative way to model personalization would be to assume that the platform can partition its customers into subgroups. For instance, suppose that the interval is partitioned into K intervals I k = [θ k 1, θ k ), with θ = θ 0 < θ 1 <... < θ K = θ. The platform only observes which interval I k each consumer belongs to. Then for each interval the platform chooses a λ k. For each interval I k, the analysis of consumers behaviour as well as of the platform s newsfeed design mirrors that of subsections 4.2 and 4.3. As K goes to infinity, we would converge to perfect personalization. 5.2 Single-homing consumers An important modeling choice that we make in this paper is to assume that the only source of friction is that users cannot choose what content they consume while on the platform. In particular, we ignore another potential source of friction, namely the existence of switching costs between websites, which could deter consumers from consuming their optimal mix of 14

15 Visit platform Visit newspaper 0 λ 1+λ 2 1 Individual demand for news ˆx(θ, q) Figure 2: Consumer decision under singlehoming content. In this subsection we test the robustness of our results by assuming that consumers incur large switching costs and are constrained to visit only one website (i.e. to singlehome). If the platform offers a share λ of news, a consumer has a choice between consuming a mix (x, y) = (λ, 1 λ) on the platform and a mix (1, 0) on the newspaper s website. The platform therefore attracts all the consumers of type θ such that U(λ, 1 λ, θ, q) U(1, 0, θ, q). To analyze this model we make some further assumptions: we assume that θ is uniformly distributed on [0, 1], that ˆx(θ, q) = α(θ + q) + βθq (if between 0 and 1), and that the utility function takes the form u(x, θ, q) = v( x ˆx(θ, q) ), where v is a non-increasing function of the difference between actual and desired consumption of news. 13 We refer to this as the linear-uniform model. For a given λ, consumers who choose to use the platform are such that 1 ˆx(θ, q) > λ ˆx(θ, q), i.e. such that ˆx(θ, q) < 1+λ (see Figure 2). The total time spent on the 2 platform is then T0 SH (q, λ) = 1+λ 2αq. But this is precisely the time spent on the platform 2(α+βq) when consumers can multihome at no cost (see Equation (8)). Given that the platform s profit is (1 λ(1 φ))t SH 0 (q, λ), the optimal λ will also be the same as in the multihoming case. A similar reasoning applies to T 1 and the optimal q. Proposition 5. In the linear-uniform model, the equilibrium values of λ, q and of firms profits are the same when consumers singlehome as when they multihome. From Proposition 5, we can conclude that the effect of content bundling by the platform on the equilibrium choice of quality as well as on the newspaper s profit is the same as in the baseline case of multihoming (Section 4). However the implications for consumer surplus are quite different. In the baseline model, content bundling does not benefit consumers, who would have been able to consume their optimal mix by multihoming. Here on the other hand, it allows platform users to consume a positive amount of both binds of content, whereas in the absence of content bundling they would have to choose a unique kind of content. 13 Given that y = 1 x in equilibrium, we drop y from the utility function. 15

16 The equivalence between multihoming and singlehoming is of course a special feature of the linear-uniform model. Yet we believe that most of the effects that we highlighted through our baseline model would carry over: the trade-off governing the choice of λ by the platform (attracting more attention by serving less profitable content), the softening of competition due to indirect traffic revenues for the newspaper, the ambiguous sign of 2 T 1, as well as the negative effect on newspaper s profit. q λ 5.3 Newspaper opt-out In practice, a newspaper with sufficient resources has the ability to remove its content from social platforms, or at least to make it harder for the platforms to show news. Given the adverse effect of content bundling on the newspaper s profit, here, we investigate how the ability to opt-out affects the equilibrium outcome. Consider the following extension of our baseline model: at τ = 0, the platform offers a contract of the form (F, φ) to the newspaper. F is a fixed payment, and φ is the share of the advertising revenue that the platform keeps whenever it shows some news to its consumers. 14 At τ = 1 the newspaper accepts or rejects the contract, and chooses a quality q. At τ = 2 the platform chooses λ if the newspaper has not opted-out. λ = 0 otherwise. At τ = 3 consumers observe q and λ and optimally allocate their attention among the two websites. Starting from τ = 1, the game is the same as in our baseline model. In particular, if the newspaper rejects the contract, its profit is π 1. To be accepted, the contract must then deliver a payoff at least equal to π 1 to the newspaper. Of course the platform does not need to offer more, and so in equilibrium the newspaper is indifferent between accepting and rejecting the offer. The platform s profit is then equal to the industry profit minus π 1. At τ = 0, the platform therefore chooses φ so as to maximize the industry profit. Because the industry revenue is constant and equal to one, the profit is maximized when the cost - i.e. the quality - is minimized. One way to do so is to offer φ = 1, i.e. to not share revenue with the newspaper. Indeed in that case, at τ = 2, the platform finds it optimal to choose λ = ˆx(θ, q) i.e. the highest desired news consumption for a quality q in the population, because by doing so it ensures that consumers spend all their time on its website (no consumer wants more news than what the platform offers). Unlike when φ < 1, there is no cost for the platform associated with showing news, because it keeps all the revenue. The newspaper then anticipates that it will get no direct traffic no matter its 14 Absent the fixed payment the newspaper would always reject the offer, as per Proposition 3. This simple two-part tariffs is actually enough to maximize profit, so there is no need to study more involved schemes (e.g. contracts dependent on q). 16

17 quality choice, and therefore chooses to not invest in quality. Proposition 6. When the platform offers a contract and the newspaper can opt-out, equilibrium quality of news is minimal. In the next section we consider a model with multiple publishers. Among other things, this will allow us to show that newspapers ability to opt-out is less critical in that context, because the platform can rely on a prisoner s dilemma logic and ensure newspapers participation without having to offer fixed payments. 6 Multiple publishers The assumption that the newspaper is the unique producer of news is clearly not innocuous. Indeed, it drives to a certain extent the softening effect : all the news consumed through the platform come from the monopolist newspaper, who therefore has less of an incentive to compete with the platform for direct traffic. In this section, we relax this assumption and study two models of competition between newspapers. The first is a model of monopolistic competition where consumers are exposed to other newspapers while on the platform. The second is a duopoly model, where there is direct competition between newspapers. 6.1 Monopolistic competition Suppose that there is a continuum of symmetric newspapers on the market. Each newspaper has a mass one of traditional readers, who decide how to allocate their time between the newspaper and the social platform. When a traditional reader of newspaper i visits the platform, he is exposed to UGC and news, in proportions 1 λ and λ. Due to the atomistic nature of the market, we assume that the news a consumer is exposed to while on the platform comes from different outlets than his usual newspaper. Newspapers are local monopolists in the sense that consumers cannot reach other newspapers directly. Advertising revenues are the same as in the baseline model. In particular the platform captures a share φ of revenues when it displays news. The timing is the following: at τ = 1 newspapers simultaneously choose their quality q, at a cost c(q). The quality of a newspaper is observed by the platform and by its traditional readers. At τ = 2 the platform chooses the share of news it displays, λ. Consumers observe λ. At τ = 3 consumers decide how to allocate their time between their usual newspaper and the platform. We look for a perfect Bayesian equilibrium where firms play a symmetric strategy and consumers form rational expectations about newspapers quality 17

18 choice (aside from their usual one, which they observe). For tractability, we assume that c(q) = cq 2, that θ is uniformly distributed on [0, 1] and we specify the following preferences: when a consumer reads a quantity x of news of quality q he gets a utility qx. Consuming an amount y of UGC gives the consumer a utility of 1 (1 θ y)2 2. With these preferences, and assuming all newspapers have the same quality q, if the consumer could choose directly which content to consume, he would spend a share ˆx(θ, q) = min{θ + q, 1} of his time reading news, and ŷ(θ, q) = 1 ˆx(θ, q) reading UGC. Benchmark (no content bundling): Without content bundling consumers have the choice between news from their usual newspaper and UGC from the platform. Each consumer then spends a share ˆx(θ, q) of his time reading news. The situation is the same as in the baseline model with a single newspaper: A newspaper s profit is π 1 = 1 0 ˆx(θ, q)dθ cq 2 = 1 q 0 (θ + q) dθ q dθ cq 2 = 1 + 2q (1 + 2c)q2. 2 The equilibrium quality is then q = 1 1+2c. Content bundling: At τ = 3, suppose that a consumer s usual newspaper has quality q and that the news quality he expects to obtain while on the platform is q. The consumer then chooses the time he spends on the platform, t, so as to maximize (1 t)q + tλq + 1 (1 θ (1 λ)t)2. The solution to this maximization problem 2 is t(θ, λ, q, q ) = max{min{ 1 q θ (1 q θ)λ, 1}, 0}. Let ˆθ (1 λ) 2 1 (λ, q, q ) be the largest solution to t(θ, λ, q, q ) = 1, and ˆθ 2 (λ, q, q ) the smallest solution to t(θ, λ, q, q ) = 0. At τ = 2, the platform chooses λ to maximize its profit. Because newspapers are atomistic, λ does not depend on a single newspaper s decision. If all newspapers except a finite number play q, the platform receives a total amount of attention T 0 (λ, q ) = 1+λ 2q Its profit is then maximized by setting λ(q ) = min{q + φ, 1}. 2 φ writes At τ = 1, suppose that newspaper i expects all other newspapers to play q. Its profit π 1 = min{ˆθ2 (λ,q i,q ),1} max{ˆθ 1 (λ,q i,q ),0} (1 t(θ, λ, q i, q )) dθ + [ max{ˆθ1 (λ,q,q ),0} + λ(1 φ) dθ T 0 (λ, q ) is given by Equation 8 with α = 1 and β = 0. min{ˆθ 2 (λ,q i,q ),1} min{ˆθ2 (λ,q,q ),1} max{ˆθ 1 (λ,q,q ),0} dθ t(θ, λ, q, q )dθ ] c(q). (10) 18

19 The first two integrals represent direct traffic to the newspaper, i.e. traffic from its usual readers, who actually observe the choice q i. The first integral is traffic by the usual readers who also visit the platform, while the second corresponds to usual readers who do not. The term between brackets correspond to indirect traffic, i.e. consumers who access the newspaper through the platform: the third integral corresponds to consumers who only visit the platform, while the fourth one represents consumers who also spend time on their favorite newspaper. Importantly, these consumers do not observe the actual q chosen by the newspaper, but rather form an expectation over the quality of news they expect to receive on the platform q, so that indirect traffic is not sensitive to q i. Assuming that c is large enough, the newspaper s profit is concave. In a symmetric configuration, we have ˆθ 1 (λ, q, q ) = λ q 0 and ˆθ 2 (λ, q, q ) = 1 q < 1. The first-order condition for a symmetric equilibrium then writes ˆθ2 (λ(q ),q,q ) ˆθ 1 (λ(q ),q,q ) t(θ, λ(q ), q, q ) 1 q dθ c (q ) = 0 q λ(q ) q Comparing q and q, we have the following result: dθ (1 λ(q )) 2 = 2cq q = 1 2c(1 λ(q )). (11) Proposition 7. In the model with monopolistic competition with additive preferences, equilibrium quality is higher with content bundling. Newspapers profits are lower. Remember that in the baseline model with a single newspaper and additive preferences content bundling lowers equilibrium quality. The intuition for the reversal of the result in a model with monopolistic competition is as follows. First, content bundling no longer creates a softening effect: when a consumer reduces the time he spends on newspaper i s website and increases the time he spends on the platform, newspaper i does not get any indirect revenue from that consumer. Therefore, the cost for a newspaper of losing direct traffic is the same with and without content bundling. Second, with content bundling, direct traffic to newspaper i is more sensitive to q i under monopolistic competition than under monopoly. Indeed, under monopoly, investment in quality by the newspaper also increases the quality of news that consumers get while on the platform. Under competition on the other hand, an increase in q i makes newspaper i more attractive without changing the value consumers expect to get from the platform. Formally, we have t(λ,θ,q i,q ) q i competition = 1 > 1 = t(λ,θ,q) (1 λ) 2 1 λ q monopoly. An interesting difference with the model with a single newspaper has to do with the possibility for newspapers to opt-out of the platform. Even though newspapers profit is lower with content bundling (by a similar argument as under monopoly), newspapers face 19

20 a prisoner s dilemma: opting-out of the platform leads a newspaper to lose indirect traffic from consumers who would not have read it otherwise, and does not allow to increase direct traffic from its usual readers. It is therefore not a viable strategy for newspapers. To achieve tractability, we have made two strong assumptions in particular. The first is that indirect traffic to newspaper i does not depend on q i, i.e. that the platform grants equal prominence to newspapers irrespective of their quality. If the platform were to favor high quality newspapers, then this would increase incentives to invest even further, reinforcing Proposition 7. The second is that newspapers do not compete head-to-head: an increase in q i does not reduce newspaper j s direct traffic. We now turn to a model of duopoly which allows us to drop these two assumptions. 6.2 Competition between duopolist newspapers Assume that there are two newspapers, i {1, 2}, of quality q i, and a platform which shows a mix of UGC and news content taken from the two newspapers. Consumers have a taste for diversity for news content, and can multi-home between the three websites. We make two further simplifying assumptions: (i) all consumers have the same preferences between the two newspapers, (ii) news quality only affects the relative preferences between the newspapers, not the relative preferences between news and UGC. Formally, these assumptions translate into the following: a consumer of type θ would like to allocate a share θ of his attention to news content, of which s i (q i, q j )θ units of attention should go to newspaper i, where we assume that s i is increasing in q i and decreasing in q j, such that s i (q i, q j ) c(q i ) is concave in q i. In this context, whatever value of λ the platform chooses, a share s i of this should come from newspaper i. Indeed, for a given λ, if the platform were to show λσ i > λs i news from i, it could increase the time spent on the platform by slightly decreasing σ i without changing its per-unit-of-attention revenue. Consumers allocation of attention: Given that news showed on the platform reflect the desired shares s 1 and s 2, consumers regard news on the platform as a perfect substitute to news obtained from direct traffic. Consumers such that θ λ would like to consume more UGC than what the platform is offering. Therefore they allocate all their attention to the platform, t (θ, λ) = 1, and end up consuming too much news. Consumers with θ > λ optimally allocate t (θ, λ) = 1 θ attention to the platform and consume their ideal 1 λ mix. The total traffic to the platform is independent of q 1 and q 2 and writes T 0 (λ) = t (θ, λ)df (θ). (12) 20

21 Platform s newsfeed design: From (12), one sees that the optimal λ does not depend on q 1 and q 2. Let λ be the solution to max λ T 0 (λ)(1 λ + λφ). Newspapers setting quality: Given that λ is independent of q 1 and q 2, newspapers take it as given when choosing their quality. Let T N (λ) = 1 T 0 (λ) be the total time spent on newspapers websites. The profit of newspaper i writes π i (q i, q j, λ) = s i (q i, q j ) (T N (λ) + (1 φ)λt 0 (λ)) c(q i ). (13) Let q (λ) be the symmetric equilibrium quality when the platform chooses λ, and R N (λ) T N (λ) + (1 φ)λt 0 (λ) the total revenue of the news industry. Proposition 8. Compared to the benchmark with λ = 0, (i) equilibrium news quality decreases when the platform shows news: q (λ ) q (0); (ii) newspapers profits also decrease. Proof. Part (i) of Proposition 8 follows from the following two lemmas. Lemma 2. q (λ) q (λ ) if and only if R N (λ) R N (λ ). Proof of Lemma 2: The best-response of newspaper i is given by s i(q i,q j ) q i R N (λ) c (q i ) = 0. One can readily check that dq i /dr N (λ) > 0, i.e. an increase in R N (λ) shifts out the reaction functions. Because, by symmetry, the reaction functions intersect on the 45 o line, the equilibrium quality is necessarily higher. 16 Q.E.D. Lemma 3. R N (λ ) R N (0). Proof of Lemma 3: The platform s profit is π(λ) = T 0 (λ)(1 λ(1 φ)) = 1 R N (λ). By revealed preferences, π(λ ) π(0), i.e. R N (λ ) R N (0). Q.E.D. To show part (ii) of the proposition, note that R N (λ ) R N (0) implies that: π i (q (λ ), q (λ ), λ ) = s i (q (λ), q (λ))r N (λ ) c(q (λ )) s i (q (λ), q (λ))r N (0) c(q (λ )). Because q (λ) q (0) and s i is decreasing in q j, we also have s i (q (λ), q (λ))r N (0) c(q (λ )) s i (q (λ), q (0))R N (0) c(q (λ )) = π i (q(λ ), q (0), 0) π i (q (0), q (0), 0), the last inequality a consequence of the fact that q (0) = argmax qi π i (q i, q (0), 0). Notice that Proposition 8 is a comparison of the benchmark (λ = 0) to the case where the platform behaves optimally (λ = λ ). In particular, it does not say that any increase 16 Absent symmetry, a shift out of the reaction functions does not guarantee that equilibrium quality increases, unless the reaction functions are upward sloping. We checked a number of specifications with asymmetric newspapers and downward sloping reaction functions, and found that equilibrium qualities increase in R(λ). 21